Mining specific and general features in both positive and negative relevance feedback

Abstract

User relevance feedback is usually utilized by Web systems to interpret user information needs and retrieve effective results for users. However, how to discover useful knowledge in user relevance feedback and how to wisely use the discovered knowledge are two critical problems. However, understanding what makes an individual document good or bad for feedback can lead to the solution of the previous problem. In TREC 2010, we participated in the Relevance Feedback Track and experimented two models for extracting pseudo-relevance feedback to improve the ranking of retrieved documents. The first one, the main run, was a pattern-based model, whereas the second one, the optional run, was a term-based model. The two models consisted of two stages: one using relevance feedback provided by TREC’10 to expand queries to extract pseudo-relevance feedback; one using pseudo-relevance feedback to find useful patterns and terms according to their relevance and irrelevance judgements to rank documents. In this paper, the detailed description of those models is presented.